fix(tests): use native_io_path for pending attachment reads on Windows

[QinLuza]

Problem

Three pending-attachment RPC tests in tests/test_gateway/test_turn_ingress_rpc.py fail on Windows with FileNotFoundError even though the material file exists on disk:

• test_pending_attachment_survives_restart_dispatches_once_and_cleans_owner
• test_pending_attachment_cancel_removes_only_its_private_owner
• test_session_delete_reclaims_pending_attachment_owner

The staged owner paths are ~297 characters (media/transcripts/<session-uuid>/.pending-chat-inputs/<64-hex owner digest>/<64-hex content sha>), past MAX_PATH (260). The production writer goes through native_io_path() (extended-length \\?\ prefix) and succeeds, but the test assertions read the bare pathlib object, which the Win32 layer refuses — so the assert sees "no such file" while the file is actually there.

Live reproduction confirmed this: Test-Path on the bare path returned False while the same path with the \\?\ prefix returned True. The DB-level assertions (which never touch disk) pass; only the bare read_bytes() asserts fail.

Fix

Route the four affected test assertions through native_io_path(), matching the convention documented in paths.py ("use this value only at the filesystem or SQLite boundary"). No product code changes.

Verification

• Test file: 71 passed on Windows (previously 3 failed / 68 passed)
• ruff: clean
• Pre-existing upstream bug: the same 3 failures reproduce on clean main in an isolated worktree, unrelated to any in-flight PR.